Abstract
Bayesian calibration of process-based models requires a prior distribution for each model parameter. Despite decades of methodological progress, researchers almost always default to uniform priors, largely because constructing informative priors from scientific literature is time-consuming and demands both domain expertise and statistical proficiency.
We introduce Distribird, an agentic web application that automates this process. Given a parameter name, physical description, and domain context, Distribird orchestrates a multi-agent pipeline that searches the scientific literature, extracts and weights reported values by domain relevance, and fits a probability distribution using AIC model selection. When no relevant literature is available, the system falls back to sensible uninformative alternatives, while clearly reporting both the evidence base and the confidence level for every prior it generates.
The tool is designed for problems where model parameters are physically interpretable and domain knowledge exists in published literature. We evaluate Distribird on 24 parameters across 10 scientific domains, comparing three open-weight models (Qwen3.6 27B, Gemma 4 31B, and Mistral Small 4 119B) against a single-prompt LLM baseline. On prior quality, the full pipeline matches the baseline. Crucially, every prior is traced to the specific papers and values from which it was constructed, and a built-in validity layer declines to produce priors for out-of-scope requests—while the single-prompt baseline returns confident but unfounded priors for 11 of 30 model–parameter cases. Furthermore, all language-model calls run locally, ensuring that no parameter descriptions or unpublished modelling details are transmitted to third-party LLM providers (only generated search terms reach public literature databases).
For scientific use, we argue that these properties matter more than marginal improvements in point-estimate accuracy.
via ArXiv AI